feat(sn60): fast dual-audit miner (jimcody1995-20260711-01)

[jimcody1995]

Summary

New submission jimcody1995-20260711-01 built to beat PR #124 (Daedalus-Icarus, current king — 26 TP / 22% detection).

Why PR #106 lost

Issue	What happened
Execution screener timeout	Agent exceeded 180s on sherlock_oku_2024_12 → kata:invalid, never scored
Round result	0 TP vs king 5 TP (when re-run completed)

Root cause: 220s runtime budget + slow rglob discovery + 46k-char deep prompts.

Strategy vs king PR #124

King (PR #124)	Our agent
Structural local patterns	Generic structural probes (receive auto-stake, domain replay, ACL gaps, reentrancy)
Map triage + findings	Same: call 1 repo map
4-file + 5-file dual audits	3-file + 4-file dual audits (critical-path + cross-contract)
.sol / .vy / .cairo	Same multi-language support
26×60 runtime (scorer-side)	165s cap — finishes under 180s screener
rglob discovery	os.walk with pruned dirs + early file cap

3-call budget

1. Call 1 — repository map triage (targets + strong map findings)
2. Call 2 — deep audit top 3 triage-ranked files
3. Call 3 — deep audit diverse second batch (4 files, different dirs)

Zero-call structural probes run before any LLM call.

Self-check

• uv run kata submission validate — pass (screening: pass)
• agent_main() smoke-safe (no crash, valid JSON)
• Not AST-equivalent to king (similarity ~0.54)
• RUN_CAP = 165s (under 180s screener)
• Await next round

Made with Cursor